# Skywork-R1V

> Source: https://aiwiki.ai/wiki/skywork_r1v
> Updated: 2026-10-13
> Categories: AI Models, Chinese AI, Large Language Models, Multimodal AI, Reasoning Models
> License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/) - attribute to "AI Wiki (aiwiki.ai)"
> Cite as: AI Wiki. "Skywork-R1V." aiwiki.ai, 13 Oct 2026. https://aiwiki.ai/wiki/skywork_r1v
> From AI Wiki (https://aiwiki.ai), the free encyclopedia of artificial intelligence. Reuse freely with attribution.

**Skywork-R1V** (also written **Skywork R1V**) is a family of open-weight [multimodal](https://aiwiki.ai/wiki/multimodal_ai) [vision-language models](https://aiwiki.ai/wiki/vision-language_model) built for chain-of-thought reasoning, developed by Skywork AI, the AGI and AIGC division of the Chinese technology company [Kunlun Tech](https://aiwiki.ai/wiki/kunlun_tech) (Beijing Kunlun Tech Co., Ltd., or Kunlun Wanwei).[1][2] The series brings the long, step-by-step "thinking" style popularized by text [reasoning models](https://aiwiki.ai/wiki/reasoning_models) such as [DeepSeek-R1](https://aiwiki.ai/wiki/deepseek_r1) to inputs that combine images and text, letting the model reason through math and science diagrams, charts, and other visual problems before answering. The first model, Skywork-R1V-38B, was open-sourced on March 18, 2025; it coupled an InternViT vision encoder with a [DeepSeek-R1-Distill](https://aiwiki.ai/wiki/deepseek_r1_distill) language backbone through a lightweight projector, and Skywork promoted it as the industry's first open-source multimodal reasoning model with visual [chain-of-thought](https://aiwiki.ai/wiki/chain_of_thought) capabilities.[1][3][2][4] Two more open generations followed in 2025, Skywork-R1V2 in April and Skywork-R1V3 in July, each a 38-billion-parameter model published with weights, inference code, and a technical report on [Hugging Face](https://aiwiki.ai/wiki/hugging_face), GitHub, and arXiv under the [MIT License](https://aiwiki.ai/wiki/mit_license), with AWQ and [GGUF](https://aiwiki.ai/wiki/gguf) quantizations also released.[5][6][7][8][9][10] A fourth, closed-source agentic variant, Skywork-R1V4-Lite, followed in November 2025.[11]

The family name appears in both forms in Skywork's own materials: the GitHub repository, the Hugging Face model cards, and the R1V3 report use the hyphenated "Skywork-R1V", while the arXiv titles of the first two reports write "Skywork R1V".[1][5][7]

## Overview

The R1V models address a gap that opened in early 2025: text-only reasoning models like DeepSeek-R1 and OpenAI's o1 showed that long reinforcement-learned chains of thought could dramatically improve performance on math and code, but those gains did not automatically extend to problems posed as images. Skywork's stated goal was to transfer this reasoning ability into the multimodal setting efficiently, without retraining a large language backbone from scratch, so that a model could examine a figure or diagram and reason about it in the same deliberate, multi-step way.[5][12]

Each open generation is a 38B-parameter model that pairs a vision encoder with a reasoning-capable language backbone, connected by a lightweight projector or "connector" module, and is post-trained with [reinforcement learning](https://aiwiki.ai/wiki/reinforcement_learning) tailored to reasoning. Skywork releases the models under the permissive MIT license and reports benchmark results positioning them among the strongest open multimodal reasoners; as with all vendor-reported figures, the numbers in this article are Skywork's own claims.[6][7][3]

## Background

### Kunlun Tech and Skywork AI

Beijing Kunlun Tech Co., Ltd. (Kunlun Wanwei) is a Beijing-based internet company founded in 2008 by Zhou Yahui and listed on the Shenzhen Stock Exchange since 2015 (ticker SZ:300418). It operates businesses in distribution, social networking, games, and, since the early 2020s, [generative AI](https://aiwiki.ai/wiki/generative_ai).[13] In June 2023, the company launched the "Tiangong" / "Skywork" large [language model](https://aiwiki.ai/wiki/language_model) brand and was included on China's "Next Tens of Billions of AIGC Products" list.[13] In October 2023 it open-sourced the Skywork-13B bilingual (Chinese and English) [foundation model](https://aiwiki.ai/wiki/foundation_models), trained on 3.2 trillion tokens and released under the Skywork Community License together with the 150B-token SkyPile Chinese corpus.[14][15] In 2024 it followed with Skywork-MoE, a 146B-parameter [mixture-of-experts](https://aiwiki.ai/wiki/mixture_of_experts) model with roughly 22B activated parameters that was upcycled from dense Skywork-13B checkpoints.[16][17] Through later integration, the AGI and AIGC business was consolidated under the Skywork AI subsidiary, which also produces the SkyReels video models, the SkyMusic / Mureka music platform, and the Skyo real-time voice assistant.[13][18]

The R1V series sits alongside Skywork's text reasoning line. In November 2024 the team unveiled Skywork o1, an early Chinese reasoning model, and in April and May 2025 it open-sourced the Skywork-OR1 (Open Reasoner 1) series of math- and code-focused 7B and 32B models, whose performance Skywork reported as approaching DeepSeek-R1 on reasoning tasks.[19][12][20] Where OR1 targets text, R1V extends the same reasoning philosophy to vision-language inputs.

### From text reasoning to multimodal reasoning

The release of DeepSeek-R1 in January 2025 and its companion DeepSeek-R1-Distill series demonstrated that long chain-of-thought reasoning, trained primarily with rule-based reinforcement learning, could transfer to dense student models in the 1.5B to 70B parameter range.[21] Skywork-R1V was conceived as the multimodal extension of that paradigm: rather than retraining a vision-language model from scratch, Skywork's researchers attached an existing vision tower to a reasoning-distilled language model and trained the connecting modules and the language model's vision-conditioned behavior.[5][3] The team frames this as an "efficient multimodal transfer method" that preserves the textual reasoning of the R1-series LLM while granting it the ability to read images.[5]

## Skywork-R1V (March 2025)

Skywork-R1V-38B was released on March 18, 2025 via the SkyworkAI GitHub organization and the `Skywork/Skywork-R1V-38B` repository on Hugging Face, with a follow-on AWQ quantized release on March 26, 2025 enabling single-GPU inference on accelerators with more than 30 GB of memory (the AWQ model card recommends 40 GB or more for long contexts).[1][3][22] The technical report, *Skywork R1V: Pioneering Multimodal Reasoning with Chain-of-Thought* by Yi Peng and colleagues, was posted to arXiv as 2504.05599 on April 8, 2025.[5]

### Architecture

The model is a modular [vision-language model](https://aiwiki.ai/wiki/vision_language_model) in which a vision tower, a learned multilayer perceptron (MLP) adapter, and a reasoning-capable [LLM](https://aiwiki.ai/wiki/llm) are wired in series.[3][5] Specifically:

- The visual backbone is **InternViT-6B-448px-V2_5**, a roughly 6-billion-parameter [Vision Transformer](https://aiwiki.ai/wiki/vision_transformer) derived from Shanghai AI Lab's [InternVL](https://aiwiki.ai/wiki/internvl) family that ingests 448-pixel image tiles.[3][8]
- The language backbone is **DeepSeek-R1-Distill-Qwen-32B**, the 32B DeepSeek-R1-Distill checkpoint built on the [Qwen](https://aiwiki.ai/wiki/qwen) 2.5 architecture.[3][5]
- A lightweight MLP **visual projector** maps the vision encoder's output space into the language model's input space; this projector is the only component initialized from scratch.[5]

Adding the 6B vision tower to the 32B language model gives an aggregate parameter count of approximately 38B, which the team uses as the canonical model size.[3] Training used a 16,384-token context length.[5]

### Training pipeline

The R1V technical report describes a three-stage training recipe designed to graft visual grounding onto an already-reasoning-capable LLM without disturbing its text-only reasoning quality: [5]

1. **MLP initialization (alignment proxy).** With the vision encoder and language model frozen, the MLP projector is first aligned to a non-reasoning substitute language model, **Qwen2.5-32B-Instruct**, on image-text data. This step, run in successive passes over 2 million, 200,000, and 40,000 samples with learning rates of 2 x 10^-4 and then 4 x 10^-5 and a batch size of 512, trains the projector to translate vision tokens into a space the Qwen family can read.[5]
2. **Model re-assembly.** The trained MLP is then transferred and used to splice the same vision encoder to the reasoning-distilled DeepSeek-R1-Distill-Qwen-32B backbone. Because R1-Distill and Qwen2.5-Instruct share the same Qwen 2.5 architecture, the projected vision tokens slot in without requiring full retraining of either tower.[5]
3. **Hybrid optimization.** The assembled model is trained for visual reasoning with a hybrid loop combining iterative [supervised fine-tuning](https://aiwiki.ai/wiki/supervised_fine-tuning) (four iterations of SFT on curated multimodal reasoning data, filtered by a reward model with a rising quality threshold, at a learning rate of 1 x 10^-4 for the first iteration and 2 x 10^-5 thereafter) and [Group Relative Policy Optimization (GRPO)](https://aiwiki.ai/wiki/grpo), the same reinforcement learning algorithm used for DeepSeek-R1, run at a learning rate of 1 x 10^-6.[5]

Layered on top of these stages is an **adaptive-length chain-of-thought distillation** procedure that dynamically adjusts the length of the reasoning trace to avoid "overthinking" on simple visual questions while preserving long traces for hard ones.[4][5] Together, these are the two contributions the report highlights in its abstract: a hybrid SFT plus GRPO optimization strategy and adaptive-length CoT distillation.[5]

### Benchmark results

Skywork reports the following results for Skywork-R1V-38B, mixing standard multimodal benchmarks with text-only reasoning benchmarks to verify that the vision grafting does not degrade the underlying R1-distill reasoning ability.[3][5]

| Benchmark | Skywork-R1V-38B |
|---|---|
| [MMMU](https://aiwiki.ai/wiki/mmmu) (val) | 69.0[3][5] |
| [MathVista](https://aiwiki.ai/wiki/mathvista) (mini) | 67.5[3][5] |
| [GPQA](https://aiwiki.ai/wiki/gpqa) (pass@1) | 61.6[3] |
| [MATH-500](https://aiwiki.ai/wiki/math_500) (pass@1) | 94.0[3][5] |
| [AIME 2024](https://aiwiki.ai/wiki/aime_2024) (pass@1) | 72.0[3][5] |

Skywork's own write-up positions the MMMU and MathVista numbers as comparable to closed-source reasoning systems such as [Gemini 2.0](https://aiwiki.ai/wiki/gemini) and [Kimi K1.5](https://aiwiki.ai/wiki/kimi_k1_5) for visual question answering, while the MATH-500 and AIME 2024 numbers are intended to show that text-only reasoning is preserved relative to the underlying DeepSeek-R1-Distill-Qwen-32B checkpoint.[2][3][4] The Hugging Face paper page reports the same headline numbers.[23]

## Skywork-R1V2 (April 2025)

Skywork-R1V2 was released on April 24, 2025, with the technical report *Skywork R1V2: Multimodal Hybrid Reinforcement Learning for Reasoning* (Peiyu Wang and colleagues, arXiv 2504.16656) posted on April 23, 2025.[8][6] An AWQ quantized release followed on April 28, 2025.[1][9] The model retains the 38B size and the InternViT-6B-448px-V2_5 vision encoder of the original R1V, but switches the language backbone from DeepSeek-R1-Distill-Qwen-32B to Alibaba's [QwQ-32B](https://aiwiki.ai/wiki/qwq) reasoning model and overhauls the training procedure.[8][6][24]

### Hybrid reinforcement learning

R1V2 frames training as a hybrid reinforcement learning problem that jointly leverages two complementary objectives: [6][25]

- **Mixed Preference Optimization (MPO)**, a preference-learning objective that combines signals from a dedicated multimodal [reward model](https://aiwiki.ai/wiki/reward_model) (called R1V-RM in the first version of the report and Skywork-VL Reward in later revisions) with rule-based constraints on format correctness, factual consistency, and step-by-step reasoning completeness.[6][25][26]
- **Group Relative Policy Optimization (GRPO)**, the same on-policy RL algorithm used in R1V, retained to drive exploration of harder reasoning trajectories by scoring each candidate answer relative to others in the same group.[6][25]

The team also introduces a **Selective Sample Buffer (SSB)**, a replay mechanism that caches high-quality training examples with non-zero advantage and reintroduces them during later policy updates. SSB is presented as a fix for the "vanishing advantages" problem in GRPO, where most rollouts in a group end up with similar rewards and contribute weak gradients.[6][25] Caching and replaying informative samples increases gradient density and, according to the report, encourages deeper chains of reasoning.[6][25]

The authors additionally report that overly strong reinforcement signals can induce visual hallucinations and discuss calibration thresholds to mitigate this trade-off between reasoning depth and visual faithfulness.[6][25] The first version of the report stated that MPO lowered the model's hallucination rate to 8.7%, compared with 12.6% for [Direct Preference Optimization](https://aiwiki.ai/wiki/direct_preference_optimization) and 18.4% for plain SFT; the revised report drops the DPO comparison, and its table lists 12.1% for SFT, 8.7% for MPO, and 9.1% for MPO combined with GRPO.[6]

### Benchmark results

The R1V2 model card and arXiv report give the following figures for Skywork-R1V2-38B, with the R1V-38B numbers that the R1V2 materials list for reference.[8][6][25]

| Benchmark | R1V2-38B | R1V-38B (as listed in R1V2 materials) |
|---|---|---|
| [MMMU](https://aiwiki.ai/wiki/mmmu) (val) | 73.6 | 68.0[8] |
| [MMMU-Pro](https://aiwiki.ai/wiki/mmmu-pro) | 52.0 | not reported[6] |
| MathVista (mini) | 74.0 | 67.0[8] |
| OlympiadBench | 62.6 | 40.4[8][6] |
| [AIME 2024](https://aiwiki.ai/wiki/aime_2024) | 78.9 | 72.0[8][6] |
| [LiveCodeBench](https://aiwiki.ai/wiki/livecodebench) | 63.6 | 57.2[8][6] |
| [GPQA](https://aiwiki.ai/wiki/gpqa) | 61.6 | not listed[8] |

The R1V row in the R1V2 materials (68.0 on MMMU and 67.0 on MathVista) differs slightly from the 69.0 and 67.5 that the original R1V report and model card give for the same model.[3][8] The GPQA figure for R1V2 appears on the Hugging Face model card rather than in the paper.[8] The most striking jumps are on OlympiadBench, where R1V2 lifts the score from 40.4 to 62.6, and on MMMU. Skywork frames the 73.6 MMMU result as the highest then-reported score for any open-source 38B-class multimodal model, and the paper states that it outperformed proprietary systems available at the time such as Claude 3.5 Sonnet (70.4) and Gemini 2 Flash (70.7).[8][25][6]

## Skywork-R1V3 (July 2025)

Skywork-R1V3-38B was open-sourced on July 9, 2025, with the model card hosted at `Skywork/Skywork-R1V3-38B`; its technical report (Wei Shen and colleagues, arXiv 2507.06167) was first posted on July 8, 2025.[1][10][7] Unlike its predecessors, R1V3 is built directly on the InternVL3-38B pretrained checkpoint, which itself combines the InternViT-6B-448px-V2_5 vision encoder with a Qwen2.5-32B language model, and concentrates its gains in post-training reinforcement learning rather than reasoning-focused pretraining.[10][27][7]

The R1V3 model card and report highlight several methodological choices: [10][7]

- a **fine-grained cold-start** supervised phase (about 20,000 samples) that prepares the model for RL, using a high-quality multimodal reasoning set built with rejection sampling and cold-start data distilled from the previous generation;
- a reinforcement-learning stage that, according to the report's abstract, raised the model's MMMU score from 64.3% to 76.0%;
- a **connector-only fine-tuning** step after RL that further boosts performance; and
- an "Entropy of Critical Reasoning Tokens" metric used to select checkpoints.

Skywork also reports that R1V3 adapts the length of its reasoning chain to input difficulty to avoid "overthinking" simple questions.[7][28]

Reported benchmarks include 76.0 on MMMU (val), 77.1 on MathVista (mini), 78.5 on the MMK12 multimodal reasoning set, 59.6 on MathVerse (vision-only), 55.4 on [MMMU-Pro](https://aiwiki.ai/wiki/mmmu-pro), and 28.5 on VisuLogic.[10][7] Skywork presented the 76.0 MMMU figure as a new open-source state of the art, approaching the human-expert reference of 76.2 and, by its comparisons, ahead of Claude 3.7 Sonnet (75.0) and GPT-4o (70.7) on that benchmark.[10][28] Such cross-vendor comparisons depend on evaluation conditions and should be treated with caution.

## Skywork-R1V4-Lite (November 2025)

Skywork-R1V4-Lite was announced on November 18, 2025. Unlike its predecessors, it is closed-source: there are no public weights, and it was offered as an API service on the Skywork platform, with the announcement saying it was "coming soon" to OpenRouter.[11][1] It is described as a lightweight, more agentic reasoning model built on Qwen3-VL-30B-A3B-Instruct (a 30B mixture-of-experts model with about 3B activated parameters), and Skywork emphasizes tool use: code execution, deep research via search-tool integration, streaming output, and multi-turn reasoning. Reported headline numbers include 91.8 on HIRbench-4K FSP and 71.4 on MME-Real Overall.[11] Because it is not open-weight, it represents a departure from the open-source releases of R1V through R1V3, although its Qwen3-VL base remains Apache 2.0 licensed.[11] As of mid-2026 the R1V4-Lite material lives on a separate branch of the Skywork-R1V GitHub repository rather than in the main README.[11]

## Multimodal reasoning approach

Across the open-weight generations, Skywork's design treats the vision encoder and the language backbone as largely fixed, capable components and focuses effort on (1) aligning them through a small connector and (2) eliciting reasoning through reinforcement learning rather than expensive multimodal pre-training. This "efficient transfer" philosophy is what lets a 38B model inherit the reasoning behavior of a strong text model while gaining the ability to see.[5][3]

The reinforcement-learning methods evolved across versions: R1V used iterative SFT plus GRPO with adaptive-length CoT distillation; R1V2 introduced the MPO-plus-GRPO hybrid with the Selective Sample Buffer; and R1V3 emphasized cold-start fine-tuning followed by RL on a curated, rejection-sampled reasoning dataset. A recurring theme is controlling reasoning length so the model spends more tokens on hard visual-math problems and fewer on easy ones.[6][7]

## Specifications

The table summarizes the open-weight R1V generations. Benchmark figures are Skywork's reported values from each model's own report and model card.

| Attribute | Skywork-R1V | Skywork-R1V2 | Skywork-R1V3 |
|---|---|---|---|
| Open-source release | March 18, 2025 | April 24, 2025 | July 9, 2025 |
| Technical report | arXiv 2504.05599 | arXiv 2504.16656 | arXiv 2507.06167 |
| Parameters | 38B | 38B | 38B |
| Vision encoder | InternViT-6B-448px-V2_5 | InternViT-6B-448px-V2_5 | InternViT-6B-448px-V2_5 (within InternVL3-38B) |
| Language backbone | DeepSeek-R1-Distill-Qwen-32B | QwQ-32B | Qwen2.5-32B (within InternVL3-38B) |
| Connector | Lightweight MLP projector | Lightweight MLP projector | MLP projector (fine-tuned alone after RL) |
| Training emphasis | Iterative SFT + GRPO, adaptive CoT distillation | Hybrid RL: MPO + GRPO + SSB | Cold-start SFT + RL post-training |
| License | MIT | MIT | MIT |
| MMMU (val) | 69.0 | 73.6 | 76.0 |
| MathVista (mini) | 67.5 | 74.0 | 77.1 |

Sources: [5][6][7][3][8][10][27]. Code and weights are distributed through the SkyworkAI/Skywork-R1V GitHub repository and the Skywork organization on Hugging Face.[1][29]

## Skywork's broader catalog

Skywork-R1V sits inside a wider open-source release program that began with the company's 2023 bilingual base model. The table below summarizes the major model families the Skywork organization has published on Hugging Face and GitHub.[29]

| Family | Year(s) | Description | Sources |
|---|---|---|---|
| Skywork-13B | 2023 | Bilingual (Chinese/English) [LLM](https://aiwiki.ai/wiki/llm) trained on 3.2T tokens; includes Base, Chat, and Math variants; released with SkyPile-150B corpus. | [14][15] |
| Skywork-MoE | 2024 | 146B-parameter [mixture-of-experts](https://aiwiki.ai/wiki/mixture_of_experts) model with 16 experts and about 22B active parameters; initialized from Skywork-13B; introduces gating logit normalization and adaptive auxiliary loss coefficients. | [30][17][16] |
| Skywork o1 and Skywork-o1-Open-PRM | November 2024 | Reasoning models (Open, Lite, and Preview) plus an open process reward model series for step-by-step reasoning supervision, based on a Qwen-2.5-1.5B backbone. | [19][29] |
| Skywork-R1V series | 2025 | InternViT plus R1-style backbone multimodal reasoning models (R1V, R1V2, R1V3, R1V4-Lite). | [1] |
| Skywork-OR1 | April / May 2025 | "Open Reasoner 1" math and code reasoning models at 7B and 32B, built on DeepSeek-R1-Distill-Qwen checkpoints and trained with large-scale rule-based reinforcement learning; previews released April 13 and final models May 13, 2025; Apache 2.0 code license. | [20][12] |
| Skywork-VL Reward | May 2025 | Multimodal [VLM](https://aiwiki.ai/wiki/vlm) reward model based on Qwen2.5-VL-7B-Instruct with a value head; reported state-of-the-art VL-RewardBench results. | [26] |
| Skywork-Reward / Reward-V2 | 2024 / July 2025 | Text reward model series; the V2 release in July 2025 includes 8 models from 0.6B to 8B parameters and was reported to top Reward Bench v1/v2, RM-Bench, JudgeBench and other reward benchmarks. | [31] |
| Skywork-UniPic, UniPic2, UniPic3 | 2025 | Unified autoregressive vision-language image generation and multi-image composition models. | [29] |
| Skywork-SWE-32B | 2025 | Software engineering reasoning model studying scaling laws for SWE tasks. | [29] |
| SkyReels-V1 / V3 / V4 | 2025 | Human-centric video foundation and multimodal video-audio generation models. | [29] |
| Mureka / Skyo | 2025 | Mureka O1 music reasoning model and the Skyo real-time voice assistant launched alongside Skywork 4.0. | [18][32] |

## Significance

Skywork-R1V is one of the earliest fully open-weight attempts to port the long-CoT, RL-trained reasoning paradigm popularized by [OpenAI o1](https://aiwiki.ai/wiki/o1) and DeepSeek-R1 into vision, and one of the more visible examples of open multimodal reasoning models emerging from Chinese labs in 2025, a wave that also includes Alibaba's QvQ, OpenGVLab's InternVL line, and Moonshot's [Kimi](https://aiwiki.ai/wiki/kimi_k2)-VL. By open-sourcing 38B models under the MIT license, Skywork made visual chain-of-thought reasoning accessible to researchers and developers who could not run or fine-tune proprietary multimodal systems.[6][7] Three aspects make it methodologically interesting:

1. **Reasoning-preserving multimodal transfer.** By aligning an MLP projector against a substitute LLM and then attaching the vision tower to a reasoning-distilled LLM before the joint SFT and GRPO loop, the team aims for "near-lossless" preservation of the underlying chain-of-thought behavior on text benchmarks like MATH-500 and AIME 2024 while still gaining visual grounding.[3][5]
2. **Hybrid RL for VLMs.** R1V2's combination of MPO with GRPO and the Selective Sample Buffer extends the GRPO recipe (originally text-only in DeepSeek) to vision-language settings, and surfaces the trade-off between reasoning depth and visual hallucination under strong reward signals. Skywork positioned these techniques as ways to raise reasoning quality while limiting hallucination.[6][25]
3. **Open weights, MIT license.** Together with the contemporaneous Skywork-OR1, Skywork-VL Reward, and Skywork-Reward-V2 releases, R1V is part of an unusually transparent open ecosystem for reasoning research, including model weights, AWQ and GGUF quantizations, training procedures, and benchmark scripts.[1][29][20][26]

The series also illustrates how quickly the open multimodal-reasoning field moved: across roughly four months in 2025, the reported MMMU figure for the family rose from 69.0 to 76.0, and the team shifted from a DeepSeek-R1-distilled backbone to QwQ-32B to the InternVL3 base. The November 2025 move to a closed-source, tool-using R1V4-Lite marked a turn toward agentic, productized deployment alongside the open research line.[7][11]

## Limitations and criticisms

The R1V technical report and the R1V2 follow-up explicitly note several limitations.[5][6][25]

- **Visual hallucinations under strong RL signals.** R1V2 reports that excessively aggressive reinforcement signals push the model toward longer but more hallucinated reasoning traces, motivating calibrated reward thresholds.[6][25]
- **Substitution dependence on a non-reasoning proxy.** R1V's projector is first aligned with Qwen2.5-32B-Instruct rather than the actual reasoning backbone, a two-step transfer that works around the difficulty of training projectors directly against an already-CoT-trained LLM.[5]
- **Hardware footprint.** At 38B parameters, the unquantized models are heavy to serve (the R1V model card's example inference script spreads the model across two GPUs); the AWQ and GGUF quantizations were released specifically to bring the family onto a single GPU with more than 30 GB of memory and onto CPU inference setups.[1][3][22]
- **Benchmark coverage gaps.** Cross-version comparisons must use only the benchmarks each report actually publishes. For R1V, the documented set in its own report and model card is MMMU, MathVista, MATH-500, AIME 2024, and GPQA; its OlympiadBench and LiveCodeBench figures appear only in the R1V2 materials, and some benchmarks named in third-party descriptions (for example, AI2D for the original R1V) are not reported by Skywork at all.[3][5][8][6]
- **Closed R1V4-Lite.** R1V4-Lite is closed-source and API-only, breaking the open-weight pattern of R1V, R1V2, and R1V3.[11]

## Related work

Skywork-R1V's design sits at the intersection of three lines of work:

- The **R1-style reasoning paradigm**: DeepSeek-R1, DeepSeek-R1-Distill, GRPO, QwQ-32B, OpenAI o1, and [Kimi K1.5](https://aiwiki.ai/wiki/kimi_k1_5), all of which use long chain-of-thought traces trained with RL on math, code, and science tasks.[21][24]
- The **InternViT / InternVL** family of open vision encoders and models from Shanghai AI Laboratory, whose InternViT-6B-448px-V2_5 encoder is used by R1V and R1V2 and is part of InternVL3-38B, the full base for R1V3.[3][10][27]
- Other contemporaneous Chinese open multimodal reasoning and vision-language models such as [QvQ](https://aiwiki.ai/wiki/qvq), [DeepSeek-VL2](https://aiwiki.ai/wiki/deepseek_vl2), [DeepSeek Janus](https://aiwiki.ai/wiki/deepseek_janus), [Qwen2.5-VL](https://aiwiki.ai/wiki/qwen2_5_vl), and [MiniCPM-V](https://aiwiki.ai/wiki/minicpm_v), which together populate the open-weights side of the multimodal AI landscape.

## See also

- [DeepSeek-R1](https://aiwiki.ai/wiki/deepseek_r1)
- [DeepSeek-R1-Distill](https://aiwiki.ai/wiki/deepseek_r1_distill)
- [Group Relative Policy Optimization (GRPO)](https://aiwiki.ai/wiki/grpo)
- [QwQ](https://aiwiki.ai/wiki/qwq)
- [Qwen2.5-VL](https://aiwiki.ai/wiki/qwen2_5_vl)
- [InternVL](https://aiwiki.ai/wiki/internvl)
- [MMMU](https://aiwiki.ai/wiki/mmmu)
- [MathVista](https://aiwiki.ai/wiki/mathvista)
- [MATH-500](https://aiwiki.ai/wiki/math_500)
- [LiveCodeBench](https://aiwiki.ai/wiki/livecodebench)
- [Reasoning models](https://aiwiki.ai/wiki/reasoning_models)
- [Knowledge Distillation](https://aiwiki.ai/wiki/knowledge_distillation)
- [Multimodal AI](https://aiwiki.ai/wiki/multimodal_ai)

## References

1. SkyworkAI, "Skywork-R1V (GitHub README)", GitHub, 2025-11-18. https://github.com/SkyworkAI/Skywork-R1V. Accessed 2026-05-21.
2. AIBase, "Kunlun Wanwei Open-Sources Skywork R1V Visual Reasoning Chain Model", AIBase, 2025-03-18. https://www.aibase.com/news/16387. Accessed 2026-05-21.
3. Skywork, "Skywork/Skywork-R1V-38B (model card)", Hugging Face, 2025-04-08. https://huggingface.co/Skywork/Skywork-R1V-38B. Accessed 2026-05-21.
4. AIBase, "Game Changer! Kunlun Wanwei's Skywork R1V Multimodal Reasoning Model Open-Sourced!", AIBase, 2025-03-19. https://www.aibase.com/news/16394. Accessed 2026-05-21.
5. Peng, Y. et al., "Skywork R1V: Pioneering Multimodal Reasoning with Chain-of-Thought", arXiv:2504.05599, 2025-04-08. https://arxiv.org/abs/2504.05599. Accessed 2026-05-21.
6. Wang, P. et al., "Skywork R1V2: Multimodal Hybrid Reinforcement Learning for Reasoning", arXiv:2504.16656, 2025-04-23. https://arxiv.org/abs/2504.16656. Accessed 2026-05-21.
7. Shen, W. et al., "Skywork-R1V3 Technical Report", arXiv:2507.06167, 2025-07-08. https://arxiv.org/abs/2507.06167. Accessed 2026-10-13.
8. Skywork, "Skywork/Skywork-R1V2-38B (model card)", Hugging Face, 2025-04-23. https://huggingface.co/Skywork/Skywork-R1V2-38B. Accessed 2026-05-21.
9. Skywork, "Skywork/Skywork-R1V2-38B-AWQ (model card)", Hugging Face, 2025-04-28. https://huggingface.co/Skywork/Skywork-R1V2-38B-AWQ. Accessed 2026-05-21.
10. Skywork, "Skywork/Skywork-R1V3-38B (model card)", Hugging Face, 2025-07-09. https://huggingface.co/Skywork/Skywork-R1V3-38B. Accessed 2026-05-21.
11. SkyworkAI, "Skywork-R1V4-Lite (README, r1v4 branch)", GitHub, 2025-11-18. https://github.com/SkyworkAI/Skywork-R1V/blob/r1v4/README.md. Accessed 2026-10-13.
12. "Kunlun Wanwei Open-Sources Skywork-OR1 Series Models: Exceptional Mathematical and Coding Capabilities." AIBase, April 2025. https://news.aibase.com/news/17080
13. Kunlun Tech, "Kunlun Tech Launched The 'SkyWork' Large Language Model And Was Selected Into The List of China's 'Next Tens of Billions of AIGC Products'", PR Newswire, 2023-06-08. https://www.prnewswire.com/news-releases/kunlun-tech-launched-the-skywork-large-language-model-and-was-selected-into-the-list-of-chinas-next-tens-of-billions-of-aigc-products-301836921.html. Accessed 2026-05-21.
14. Kunlun Tech, "Kunlun Tech releases open source 13B high-quality commercial large model, ahead of Llama2 and Baichuan2", Kunlun Tech News, 2023-10-30. http://www.kunlun.com/2023/en_mnews_1030/328.html. Accessed 2026-05-21.
15. Wei, T. et al., "Skywork: A More Open Bilingual Foundation Model", arXiv:2310.19341, 2023-10-30. https://arxiv.org/abs/2310.19341. Accessed 2026-05-21.
16. "SkyworkAI/Skywork-MoE: A Deep Dive into Training Techniques for Mixture-of-Experts Language Models." GitHub. https://github.com/SkyworkAI/Skywork-MoE
17. Wei, T. et al., "Skywork-MoE: A Deep Dive into Training Techniques for Mixture-of-Experts Language Models", arXiv:2406.06563, 2024-06-10. https://arxiv.org/abs/2406.06563. Accessed 2026-05-21.
18. TMTPost, "Kunlun Tech Launches Skywork 4.0 AI Model and Skyo Real-Time Voice Assistant", TMTPost, 2025-04-18. https://en.tmtpost.com/news/7345973. Accessed 2026-05-21.
19. "Kunlun Tech Launches Skywork o1, a Sophisticated AI Model with Advanced Reasoning Capabilities." TMTPost, 2024. https://en.tmtpost.com/news/7355212
20. SkyworkAI, "Skywork-OR1 (GitHub README)", GitHub, 2025-05-13. https://github.com/SkyworkAI/Skywork-OR1. Accessed 2026-05-21.
21. DeepSeek-AI, "DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning", arXiv:2501.12948, 2025-01-22. https://arxiv.org/abs/2501.12948. Accessed 2026-05-21.
22. Skywork, "Skywork/Skywork-R1V-38B-AWQ (model card)", Hugging Face, 2025-03-26. https://huggingface.co/Skywork/Skywork-R1V-38B-AWQ. Accessed 2026-10-13.
23. Hugging Face, "Paper page: Skywork R1V: Pioneering Multimodal Reasoning with Chain-of-Thought", Hugging Face Papers, 2025-04-08. https://huggingface.co/papers/2504.05599. Accessed 2026-05-21.
24. Qwen Team, "QwQ-32B (model card)", Hugging Face, 2025-03-06. https://huggingface.co/Qwen/QwQ-32B. Accessed 2026-05-21.
25. MarkTechPost, "Skywork AI Advances Multimodal Reasoning: Introducing Skywork R1V2 with Hybrid Reinforcement Learning", MarkTechPost, 2025-04-25. https://www.marktechpost.com/2025/04/25/skywork-ai-advances-multimodal-reasoning-introducing-skywork-r1v2-with-hybrid-reinforcement-learning/. Accessed 2026-05-21.
26. Wang, X. et al., "Skywork-VL Reward: An Effective Reward Model for Multimodal Understanding and Reasoning", arXiv:2505.07263, 2025-05-12. https://arxiv.org/abs/2505.07263. Accessed 2026-05-21.
27. OpenGVLab, "OpenGVLab/InternVL3-38B (model card)", Hugging Face, 2025-04. https://huggingface.co/OpenGVLab/InternVL3-38B. Accessed 2026-10-13.
28. "Kunlun Wanwei releases and open-sources Skywork-R1V 3.0, multimodal reasoning capability approaches human expert level." 1ai.net, July 2025. https://www.1ai.net/en/39088.html
29. Skywork, "Skywork organization page", Hugging Face, 2026-05-21. https://huggingface.co/Skywork. Accessed 2026-05-21.
30. Skywork, "Skywork/Skywork-MoE-Base (model card)", Hugging Face, 2024-06-10. https://huggingface.co/Skywork/Skywork-MoE-Base. Accessed 2026-05-21.
31. Skywork, "Skywork-Reward-V2: Leading the New Milestone for Open-Source Reward Models", PR Newswire, 2025-07-04. https://www.prnewswire.com/news-releases/skywork-reward-v2-leading-the-new-milestone-for-open-source-reward-models-302498377.html. Accessed 2026-05-21.
32. Music Business Worldwide, "China's $6B-valued Kunlun Tech debuts 'world's first' music reasoning model, claims it can outperform Suno", Music Business Worldwide, 2025-03-26. https://www.musicbusinessworldwide.com/chinas-6b-valued-kunlun-tech-debuts-worlds-first-music-reasoning-model-claims-it-can-outperform-suno/. Accessed 2026-05-21.